AI detector false-positive calculator
Inputs
Essays or other pieces run through the detector
Share of AI-written submissions the detector flags
Share of submissions written without AI that the detector flags
More options
Each one checked once
Flagged if any one of them flags
Results
- Expected flags
- 2.97 of 30
- False flags (written without AI)
- 0.27
- Chance a flagged submission was written without AI
- 9.1%
- Chance of at least one false flag
- 23.8%
About 9 of every 100 flagged submissions were written without AI.
100 flagged submissions
All numbers
- Written with AI
- 3
- Written without AI
- 27
- Correct flags (true positives)
- 2.7
- Missed (false negatives)
- 0.3
- False flags (false positives)
- 0.27
- Correctly cleared (true negatives)
- 26.7
- Positive predictive value (PPV)
- 90.9%
- Negative predictive value (NPV)
- 98.9%
- Positive likelihood ratio (LR+)
- 90
- Chance of AI use before and after a flag
- 10% → 90.9%
Published error rates
| Detector | Measure | Rate | Unit | Sample | Population and conditions | Source | Calculator |
|---|---|---|---|---|---|---|---|
| TurnitinVendor claim | False-positive rate | <1% | Document | Not given | Documents with over 20% AI writing; vendor test on 800,000 pre-ChatGPT documents | AI writing detection update from Turnitin's Chief Product Officer2023-05-23Checked 2026-10-22 | |
| TurnitinVendor claim | False-positive rate | ~4% | Sentence | Not given | Individual highlighted sentences, not whole documents | AI writing detection update from Turnitin's Chief Product Officer2023-05-23Checked 2026-10-22 | Not for whole documents |
| TurnitinVendor claim | True-positive rate | 97% | Document | Not given | ChatGPT and GPT-3 writing in Turnitin's lab, before launch | Turnitin announces AI writing detector and AI writing resource center for educators2023-02-13Checked 2026-10-22 | |
| Turnitin | False-positive rate | 0% | Document | n = 18 | 9 human-written English documents and 9 machine-translated human-written documents | Testing of detection tools for AI-generated text (Weber-Wulff et al.)2023-12-25Checked 2026-10-22 | |
| GPTZero | False-positive rate | 50% | Document | n = 18 | 9 human-written English documents and 9 machine-translated human-written documents | Testing of detection tools for AI-generated text (Weber-Wulff et al.)2023-12-25Checked 2026-10-22 | |
| Crossplag | False-positive rate | 16.7% | Document | n = 18 | 9 human-written English documents and 9 machine-translated human-written documents | Testing of detection tools for AI-generated text (Weber-Wulff et al.)2023-12-25Checked 2026-10-22 | |
| Average of 14 tools | False-positive rate | 2.4% | Document | Not given | Human-written English documents (9 per tool) | Testing of detection tools for AI-generated text (Weber-Wulff et al.)2023-12-25Checked 2026-10-22 | |
| Average of 14 tools | False-positive rate | 11.1% | Document | Not given | Human-written documents machine-translated into English (9 per tool) | Testing of detection tools for AI-generated text (Weber-Wulff et al.)2023-12-25Checked 2026-10-22 | |
| Turnitin | True-positive rate | 72.2% | Document | n = 36 | AI-generated documents, including manually edited and machine-paraphrased ones | Testing of detection tools for AI-generated text (Weber-Wulff et al.)2023-12-25Checked 2026-10-22 | |
| Average of 14 tools | False-negative rate | 19.8% | Document | Not given | Unedited ChatGPT documents, first prompt (class 03-AI, 9 per tool) | Testing of detection tools for AI-generated text (Weber-Wulff et al.)2023-12-25Checked 2026-10-22 | |
| Average of 14 tools | False-negative rate | 21.4% | Document | Not given | Unedited ChatGPT documents, second prompt (class 04-AI, 9 per tool) | Testing of detection tools for AI-generated text (Weber-Wulff et al.)2023-12-25Checked 2026-10-22 | |
| Average of 14 tools | False-negative rate | 51.6% | Document | Not given | ChatGPT documents edited by hand (synonyms, reordering; class 05-ManEd) | Testing of detection tools for AI-generated text (Weber-Wulff et al.)2023-12-25Checked 2026-10-22 | |
| Average of 14 tools | False-negative rate | 71.4% | Document | Not given | ChatGPT documents rewritten with Quillbot (class 06-Para) | Testing of detection tools for AI-generated text (Weber-Wulff et al.)2023-12-25Checked 2026-10-22 | |
| Average of 7 detectors | False-positive rate | 61.22% | Document | Not given | 91 TOEFL essays by non-native English writers | GPT detectors are biased against non-native English writers (Liang et al.)2023-07-10Checked 2026-10-22 | |
| Any of 7 detectors | False-positive rate | 97.8% | Document | n = 91 | Same 91 TOEFL essays, counted as flagged if at least one detector flagged them | GPT detectors are biased against non-native English writers (Liang et al.)2023-07-10Checked 2026-10-22 | |
| Average of 7 detectors | False-positive rate | 5.19% | Document | Not given | 88 essays by US 8th-grade students (Hewlett ASAP data) | GPT detectors are biased against non-native English writers (Liang et al.)2023-07-10Checked 2026-10-22 | |
| Average of 7 detectors | False-positive rate | 15% | Document | Not given | 10 human-written control samples; flagged at a 60% AI score | Simple techniques to bypass GenAI text detectors: implications for inclusive education (Perkins et al.)2024-09-09Checked 2026-10-22 | |
| Copyleaks | False-positive rate | 50% | Document | n = 10 | 10 human-written control samples; flagged at a 60% AI score | Simple techniques to bypass GenAI text detectors: implications for inclusive education (Perkins et al.)2024-09-09Checked 2026-10-22 | |
| Crossplag | False-positive rate | 30% | Document | n = 10 | 10 human-written control samples; flagged at a 60% AI score | Simple techniques to bypass GenAI text detectors: implications for inclusive education (Perkins et al.)2024-09-09Checked 2026-10-22 | |
| GPTZero | False-positive rate | 10% | Document | n = 10 | 10 human-written control samples; flagged at a 60% AI score | Simple techniques to bypass GenAI text detectors: implications for inclusive education (Perkins et al.)2024-09-09Checked 2026-10-22 | |
| Turnitin | False-positive rate | 0% | Document | n = 10 | 10 human-written control samples; flagged at a 60% AI score | Simple techniques to bypass GenAI text detectors: implications for inclusive education (Perkins et al.)2024-09-09Checked 2026-10-22 | |
| Average of 7 detectors | True-positive rate | 39.5% | Document | Not given | Unedited AI-generated samples (GPT-4, Claude 2, Bard); mean of binary, semi-binary and logarithmic accuracy | Simple techniques to bypass GenAI text detectors: implications for inclusive education (Perkins et al.)2024-09-09Checked 2026-10-22 | |
| Average of 7 detectors | True-positive rate | 22.14% | Document | Not given | The same AI-generated samples after six adversarial techniques; mean of three accuracy scoring approaches | Simple techniques to bypass GenAI text detectors: implications for inclusive education (Perkins et al.)2024-09-09Checked 2026-10-22 | |
| Turnitin | True-positive rate | 91% | Submission | n = 22 | 22 ChatGPT (GPT-4) submissions written with prompts meant to evade detection; flagged as containing any AI content | Detection of GPT-4 Generated Text in Higher Education: Combining Academic Judgement and Software to Identify Generative AI Tool Misuse (Perkins et al.)2023-10-31Checked 2026-10-22 | |
| Originality.ai | True-positive rate | 85% | Passage | Not given | RAID benchmark, no adversarial attack, threshold set for a 5% false-positive rate | RAID: A Shared Benchmark for Robust Evaluation of Machine-Generated Text Detectors (Dugan et al.)Checked 2026-10-22 | |
| GPTZero | True-positive rate | 66.5% | Passage | Not given | RAID benchmark, no adversarial attack, threshold set for a 5% false-positive rate | RAID: A Shared Benchmark for Robust Evaluation of Machine-Generated Text Detectors (Dugan et al.)Checked 2026-10-22 | |
| GPTZero | False-positive rate | 0% | Passage | Not given | RAID human-written text, detector score threshold 0.5 | RAID: A Shared Benchmark for Robust Evaluation of Machine-Generated Text Detectors (Dugan et al.)Checked 2026-10-22 | |
| Originality.ai | False-positive rate | 0.25% | Passage | Not given | RAID human-written text, detector score threshold 0.5 | RAID: A Shared Benchmark for Robust Evaluation of Machine-Generated Text Detectors (Dugan et al.)Checked 2026-10-22 | |
| Winston AI | False-positive rate | 0.55% | Passage | Not given | RAID human-written text, detector score threshold 0.5 | RAID: A Shared Benchmark for Robust Evaluation of Machine-Generated Text Detectors (Dugan et al.)Checked 2026-10-22 | |
| ZeroGPT | False-positive rate | 1.42% | Passage | Not given | RAID human-written text, detector score threshold 0.5 | RAID: A Shared Benchmark for Robust Evaluation of Machine-Generated Text Detectors (Dugan et al.)Checked 2026-10-22 | |
| Pangram | False-positive rate | 0.75% | Passage | Not given | Highest of six genres (restaurant reviews), at Youden-optimized thresholds; not student essays | Artificial Writing and Automated Detection (Jabarian and Imas, NBER w34223)Checked 2026-10-22 | |
| GPTZero | False-positive rate | 2.38% | Passage | Not given | Highest of six genres (Amazon reviews), at Youden-optimized thresholds; not student essays | Artificial Writing and Automated Detection (Jabarian and Imas, NBER w34223)Checked 2026-10-22 | |
| Originality.ai | False-positive rate | 2.18% | Passage | Not given | Highest of six genres (restaurant reviews), at Youden-optimized thresholds; not student essays | Artificial Writing and Automated Detection (Jabarian and Imas, NBER w34223)Checked 2026-10-22 | |
| Pangram | False-negative rate | 2% | Passage | Not given | Worst case (short Gemini 2.0 Flash restaurant reviews), at Youden-optimized thresholds; not student essays | Artificial Writing and Automated Detection (Jabarian and Imas, NBER w34223)Checked 2026-10-22 | |
| Surfer AI detector | False-positive rate | 4.2% | Document | n = 15,894 | Pre-ChatGPT web articles; article counted as AI if over 50% of its 500-word chunks were | More Articles Are Now Created by AI Than Humans (Graphite)2025-10-14Checked 2026-10-22 | |
| Surfer AI detector | False-negative rate | 0.6% | Document | n = 6,009 | Articles written with GPT-4o | More Articles Are Now Created by AI Than Humans (Graphite)2025-10-14Checked 2026-10-22 | |
| GPTZeroVendor claim | False-positive rate | 1% | Passage | Not given | Vendor's account of its RAID result, non-adversarial text | GPTZero is Officially The Most Accurate Commercial AI Detector2025-10-15Checked 2026-10-22 | |
| GPTZeroVendor claim | True-positive rate | 95.7% | Passage | Not given | Vendor's account of its RAID result, non-adversarial text, at a 1% false-positive rate | GPTZero is Officially The Most Accurate Commercial AI Detector2025-10-15Checked 2026-10-22 | |
| CopyleaksVendor claim | False-positive rate | 0.03% | Document | Not given | Vendor figure | Copyleaks AI DetectorChecked 2026-10-22 | |
| CopyleaksVendor claim | True-positive rate | 99.2% | Document | Not given | Vendor figure for English AI-written text | Copyleaks AI DetectorChecked 2026-10-22 | |
| Pangram 4Vendor claim | False-positive rate | 0.0041% | Document | Not given | Vendor's internal benchmark | Is Pangram accurate?2026-10-16Checked 2026-10-22 | |
| Pangram 4Vendor claim | False-negative rate | 0.3396% | Document | Not given | Vendor's test on nearly 520,000 AI-generated documents | Is Pangram accurate?2026-10-16Checked 2026-10-22 | |
| Pangram 4Vendor claim | False-positive rate | 0% | Document | n = 25,996 | Vendor's test on student essays from diverse backgrounds | Is Pangram accurate?2026-10-16Checked 2026-10-22 | |
| Pangram 4.1Vendor claim | False-positive rate | 0.009% | Document | n = 781,981 | Vendor's model card, English web pages (FineWeb, in-scope) | Pangram 4.1 model card2026-10-19Checked 2026-10-22 | |
| Pangram 4.1Vendor claim | False-positive rate | 0% | Document | n = 91 | Vendor's model card, the 91 Liang et al. TOEFL essays | Pangram 4.1 model card2026-10-19Checked 2026-10-22 |